The social AI hypothesis - by Stefano Viel - Ste

Training a neural network induces a policy that performs well under a particular objective and data distribution.

The current approach to improving LLMs’ capabilities is to create a sufficiently diverse set of environments, train on all of them, and hope to obtain generality. This requires continuous human contribution and, unlike pre-training, doesn’t scale. I believe that training in an environment with other agents (i.e., social training) scales with the number of agents in the environment as they create additional complexity and require more intelligence without any human intervention. Thus, social training represents a more viable path to artificial general intelligence. Don’t get me wrong, what…

saved by

• Akira Yoshiyama
• Tommy He

related reading

• On neural scaling and the quanta hypothesisericjmichaud.com
• [2510.13786] The Art of Scaling Reinforcement Learning Compute for LLMsarxiv.org
• The Scaling Hypothesis · Gwern.netgwern.net
• Will scaling work?dwarkeshpatel.com
• Interaction Models: A Scalable Approach to Human-AI Collaboration - Thinking Machines Labthinkingmachines.ai
• The Extreme Inefficiency of RL for Frontier Models - Toby Ordtobyord.com
• Scaling: The State of Play in AIoneusefulthing.org
• I. From GPT-4 to AGI: Counting the OOMs - SITUATIONAL AWARENESSsituational-awareness.ai
• Scaling is subtler than it seemsberen.io
• Scaling Laws Do Not Scalearxiv.org
• Oversight Assistants: Turning Compute into Understandingbounded-regret.ghost.io
• Solipsistic Superintelligence is Unlikely to be Cooperativearxiv.org